Flat bars are named in the record, not cut out of the calendar
A flat bar here is a daily bar with no printed range — open = high = low — or no volume. The halt auditor (M5.5) refuses a fill on one: nothing fills in a bar that did not trade, and a bar whose open, high and low are the same number carries no print an order could have met. The auditor is right. What the first survey showed is where such bars are, and what refusing them costs under the audit as implemented (ADR-0041).
The census, over the four universes' definition partitions, taken 2026-09-12 and re-taken per name for this record:
| universe | flat bars | where |
|---|---|---|
| index_etfs | 0 | — |
| sector_etfs | 3 | XLB 1998-12-29 (700 shares) and 1999-02-26; XLI 1999-01-26 (no volume). Days 5 and 45 of a calendar that starts 1998-12-22 |
| dow_30 | 9 | seven names: DIS 3 (2001-08-30 → 2002-01-16); AMGN 1994-01-04; KO 1993-04-16; CAT, CVX, PG and SHW all on 1998-05-20 |
| sp_100 | 319 | 28 names: C 254 — every bar of 1996, real volume and a different close each day, open, high and low collapsed onto the close; AMD 11 (1996); DHR 10 (1993–97); GILD 6 (three with no volume); COF 5; DIS 3; SPG 3; two each in AMT, DE, EMR, MO, TMO, UNP; single bars in sixteen names |
Three dates recur across names — 1998-05-20 in eight names, 1999-12-17 in six, 1995-01-31 in five. A day that is flat in eight large names at once is the vendor's history, not eight halts. C's 1996 is the same thing at scale: the closes are real (the year's moving averages, down-runs and returns read correctly from them), and only the prints inside each day are missing. This is a source defect in the early history of some names, recorded beside M0.3 in docs/M0-ANSWERS.md and in the Tiingo rights record's notes, not a property of the market.
What it costs. Under the sampled, whole-run audit: 2,284 cells refused and 3,598 with a refused baseline, by sort order (ADR-0041). Under ADR-0041 as proposed: one missed trade per flat bar per cell that would have entered on it, named in the record. The always-long baseline enters every box and so misses all of them — for C, 254 of its roughly 2,500 boxes in the window, one name in eighty-three.
Decision (adopted 2026-09-13)
Name the defect in the record; leave the calendars where they are.
- The sector calendar stays at 1998-12-22, the members' common span at declaration, and
sp_100's at 1993-01-04. No boundary moves for a source defect. - **C stays in
sp_100, 1996 included.** Its closes are real and its signals read from them; the fills that the year cannot give are missed trades (ADR-0041), counted by name and bar. - The bias is named where a universe is read. A superseding
UniverseDeclaredrecord forsector_etfsand forsp_100— same members, a note naming the flat bars and their count — so the universe's own record says what its history cannot give, in the same way the two constituent lists already name survivorship. Every survey cell and every verdict on these universes carries the missed census beside the trade count; a verdict is never shown without it.
This is conditional on ADR-0041. If the audit stays as it is, the flat bars refuse whole runs and the calendar question returns in the form below.
Consequences
- No apparatus change beyond ADR-0041's. Two appended universe records with notes; no calendar record.
- The sector definition partition keeps its first 45 days, and the measurement and reserve boundaries stay where record #9's cells were computed. The next grid's sector cells remain on the same partition as the first survey's.
- The missed census is part of what M12.7's programme page and M12.6's verdict evidence show; the shrinkage table is not affected.
Considered options
- **Supersede
sector_etfs's calendar to start after 1999-02-26.** Available under ADR-0038 as a recorded supersession —calendar freeze --universe sector_etfs --start 1999-03-01 --supersedes 729745-739870 --reason …— and allowed, because no registered question has used the sector cuts (a universe's cuts bind only itself). What it does: moving the start 68 days later shortens the span by 68 days; with the recorded split (0.3 / 0.5 / 0.2) the definition end moves about 48 days later and the measurement end about 14 days later — a fortnight of the sealed reserve becomes measurement, which is the drift ADR-0038 was written to stop, here by the author's act rather than the archive's growth. It removes three bars from 2,100 definition days and makes every later sector cell non-comparable to record #9's, which was computed on the earlier calendar and stays recorded on it. Not proposed; the author's to choose if the sector definition partition should be free of the defect at that price. - **Exclude C's 1996 by a superseding
sp_100record.** A universe record carries members, not per-name windows. This is either a record without C — ten clean years of one name dropped for one defective year, and its 1996 signals with them — or a new per-name exclusion field in the record and the archive slice, an apparatus change the row did not ask for. Rejected in both forms: the year's closes are real, and the audit already refuses exactly what is not. - Leave the refusals as they are and name them in every verdict. This is the status quo under the sampled audit: measurability by sort order (ADR-0041). Rejected as a standing rule; the naming it asks for is what the decision above does, with the audit fixed.